[Bug libgomp/81805] Another libgomp.c/for-5.c failure on nvptx -- illegal memory access
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81805 Tom de Vries changed: What|Removed |Added Status|RESOLVED|WAITING Last reconfirmed||2017-08-18 Resolution|FIXED |--- Ever confirmed|0 |1
[Bug libgomp/81805] Another libgomp.c/for-5.c failure on nvptx -- illegal memory access
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81805 Tom de Vries changed: What|Removed |Added Status|UNCONFIRMED |RESOLVED Resolution|--- |FIXED --- Comment #11 from Tom de Vries --- (In reply to Tom de Vries from comment #10) > [ Redoing https://gcc.gnu.org/ml/gcc-bugs/2017-08/msg01333.html, using > PR81875 instead of PR81844, since PR81844 was overwritten ] > > Filed comment 3 as PR81875 - omp for loop optimized away > > This PR remains for the analysis of the test-case from comment 0. My current hypothesis is that this is a cuda bug. Filed bug report at nvidia, waiting for confirmation.
[Bug libgomp/81805] Another libgomp.c/for-5.c failure on nvptx -- illegal memory access
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81805 --- Comment #10 from Tom de Vries --- [ Redoing https://gcc.gnu.org/ml/gcc-bugs/2017-08/msg01333.html, using PR81875 instead of PR81844, since PR81844 was overwritten ] Filed comment 3 as PR81875 - omp for loop optimized away This PR remains for the analysis of the test-case from comment 0.
[Bug libgomp/81805] Another libgomp.c/for-5.c failure on nvptx -- illegal memory access
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81805 --- Comment #10 from Tom de Vries --- Filed comment 3 as PR81844 - omp for loop optimized away This PR remains for the analysis of the test-case from comment 0.
[Bug libgomp/81805] Another libgomp.c/for-5.c failure on nvptx -- illegal memory access
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81805 Tom de Vries changed: What|Removed |Added Attachment #41970|0 |1 is obsolete|| --- Comment #9 from Tom de Vries --- Created attachment 41972 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=41972=edit Updated patch Updated patch, passes check-target-libgomp and check-gcc/gomp.exp.
[Bug libgomp/81805] Another libgomp.c/for-5.c failure on nvptx -- illegal memory access
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81805 --- Comment #8 from Tom de Vries --- (In reply to Tom de Vries from comment #7) > (In reply to Tom de Vries from comment #6) > > Created attachment 41970 [details] > > tentative patch for test-case from comment 3 > > Works for test-case from comment 2 as well. But not for test-case from comment 0.
[Bug libgomp/81805] Another libgomp.c/for-5.c failure on nvptx -- illegal memory access
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81805 --- Comment #7 from Tom de Vries --- (In reply to Tom de Vries from comment #6) > Created attachment 41970 [details] > tentative patch for test-case from comment 3 Works for test-case from comment 2 as well.
[Bug libgomp/81805] Another libgomp.c/for-5.c failure on nvptx -- illegal memory access
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81805 --- Comment #6 from Tom de Vries --- Created attachment 41970 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=41970=edit tentative patch for test-case from comment 3
[Bug libgomp/81805] Another libgomp.c/for-5.c failure on nvptx -- illegal memory access
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81805 --- Comment #5 from Tom de Vries --- (In reply to Tom de Vries from comment #4) > (gdb) call debug_generic_expr (cond) > (signed long) i < 0 And the conversion from unsigned to signed comparison is done here in c_parser_omp_for_loop: ... (gdb) call debug_generic_expr (cond) <<< Unknown tree: c_maybe_const_expr i >>> > 9223372036854775807 (gdb) n 15030 cond = c_fully_fold (cond, false, NULL); (gdb) n 15031 switch (cond_expr.original_code) (gdb) call debug_generic_expr (cond) (signed long) i < 0 ...
[Bug libgomp/81805] Another libgomp.c/for-5.c failure on nvptx -- illegal memory access
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81805
--- Comment #4 from Tom de Vries ---
(In reply to Alexander Monakov from comment #3)
> The comparison should have read '(long long)i < 0', no idea how the cast is
> lost.
It seems to be lost here, in c-omp.c:
...
553/* 2.5.1. The comparison in the condition is computed in
554 the type of DECL, otherwise the behavior is undefined.
555
556 For example:
557 long n; int i;
558 i < n;
559
560 according to ISO will be evaluated as:
561 (long)i < n;
562
563 We want to force:
564 i < (int)n; */
565if (TREE_CODE (op0) == NOP_EXPR
566&& decl == TREE_OPERAND (op0, 0))
567 {
568TREE_OPERAND (cond, 0) = TREE_OPERAND (op0, 0);
569TREE_OPERAND (cond, 1)
570 = fold_build1_loc (elocus, NOP_EXPR, TREE_TYPE
(decl),
571 TREE_OPERAND (cond, 1));
572 }
...
...
568 TREE_OPERAND (cond, 0) = TREE_OPERAND (op0, 0);
(gdb) call debug_generic_expr (cond)
(signed long) i < 0
(gdb) n
569 TREE_OPERAND (cond, 1)
(gdb) call debug_generic_expr (cond)
i < 0
...
[Bug libgomp/81805] Another libgomp.c/for-5.c failure on nvptx -- illegal memory access
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81805
--- Comment #3 from Alexander Monakov ---
The new testcase fails on any target and not related to offloading. Simplified
further:
#define N 32ULL
int a[N];
const unsigned long long c = 0x7fffULL;
f2_tpf_static32 (void)
{
unsigned long long i;
#pragma omp for
for (i = c + N; i > c; i -= 1ULL)
a[i - 1ULL - c] -= 4;
}
in the .original dump we have:
{
long long unsigned int i;
long long unsigned int i;
#pragma omp for
for (i = 9223372036854775839; i < 0; i = i + 18446744073709551615)
{
a[i + 9223372036854775808] = a[i + 9223372036854775808] + -4
}
}
and soon after we deduce that loop doesn't iterate because i is unsigned. The
comparison should have read '(long long)i < 0', no idea how the cast is lost.
[Bug libgomp/81805] Another libgomp.c/for-5.c failure on nvptx -- illegal memory access
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81805
--- Comment #2 from Tom de Vries ---
Simplified version, aborts in host code because loop has no effect:
...
#include
extern void abort ();
#define N 32ULL
#pragma omp declare target
int a[N];
#pragma omp end declare target
const unsigned long long c = 0x7fffULL;
__attribute__((noinline, noclone)) void
f2_tpf_static32 (void)
{
unsigned long long i;
#pragma omp target parallel for schedule(static, 32)
for (i = c + N; i > c; i -= 1ULL)
a[i - 1ULL - c] -= 4;
}
__attribute__((noinline, noclone)) int
test_tpf_static32 (void)
{
int i, j, k;
for (i = 0; i < N; i++)
a[i] = i - 25;
#pragma omp target update to(a)
f2_tpf_static32 ();
#pragma omp target update from(a)
for (i = 0; i < N; i++)
printf ("%d: %d, expected: %d\n", i, a[i], i - 29);
for (i = 0; i < N; i++)
if (a[i] != i - 29)
return 1;
return 0;
}
int
main ()
{
if (test_tpf_static32 ())
abort ();
return 0;
}
...
[Bug libgomp/81805] Another libgomp.c/for-5.c failure on nvptx -- illegal memory access
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81805 --- Comment #1 from Alexander Monakov --- Can't reproduce this on my end. Are you going to proceed with analyzing the failure?
